CODEWITHSUNDEEP
machine-learningdeep-learningpytorchclassificationmultilabel-classification
Sam
Sam

Reputation: 207

How to learn multiple binary classifiers?

I have an input of shape 14 x 10 x 128 x 128, where 14 is batch_size, 10 is the sequence_length and each item in the sequence is of shape 128 x 128. I want to learn to map this input to output of shape 14 x 10 x 128, i.e., for each item in the sequence I want to learn 128-binary classifiers.

Does the following model make sense? So, first I reshape my input to 140 x 128 x 128 and then pass it through the model and reshape the output back to 14 x 10 x 128.

classifier = nn.Sequential(
    nn.Conv1d(128, 128, 1),
    nn.ReLU(),
    nn.BatchNorm1d(128),
    nn.Conv1d(128, 128, 1),
    nn.ReLU(),
    nn.BatchNorm1d(128),
    nn.Conv1d(128, 1, 1)
)

Thank you.

Upvotes: 1

Views: 115

Answers (1)

Ivan
Ivan

Reputation: 40668

Not really convinced a 1D convolution will get you anywhere since it reasons in two dimensions only. In your case, you are dealing with a sequence of 2D elements.
Naturally a nn.Conv2d would seem more appropriate for this kind of task.

You are looking to do a one-to-one mapping with your sequence elements and can therefore consider each one of them as an independent instance.
Then a straightforward approach is to simply collapse the sequence into the batch axis and use a CNN coupled with a fully-connected layer.

Here is a very minimal example with a single layer:

model = nn.Sequential(nn.Conv2d(1, 8, 2),
                      nn.ReLU(),
                      nn.AdaptiveAvgPool2d(1),
                      nn.Flatten(),
                      nn.LazyLinear(128))

This requires you to reshape the tensor before and after to collapse and expand the sequence dimensions:

>>> x = torch.rand(14, 10, 128, 128)
>>> y = model(x.view(-1,1,128,128)).view(-1,10,128)

Upvotes: 1

Related Questions